U.S. patent application number 15/482856 was filed with the patent office on 2017-12-07 for ising device and control method thereof.
This patent application is currently assigned to FUJITSU LIMITED. The applicant listed for this patent is FUJITSU LIMITED. Invention is credited to SATOSHI MATSUBARA, Hirotaka TAMURA.
Application Number | 20170351949 15/482856 |
Document ID | / |
Family ID | 60483357 |
Filed Date | 2017-12-07 |
United States Patent
Application |
20170351949 |
Kind Code |
A1 |
MATSUBARA; SATOSHI ; et
al. |
December 7, 2017 |
ISING DEVICE AND CONTROL METHOD THEREOF
Abstract
An individual neuron circuit calculates a first value based on a
sum of products each obtained by multiplying one of weight values,
each representing connection or disconnection between a
corresponding neuron circuit and one of the other neuron circuits,
by a corresponding one of output signals of the other neuron
circuits and outputs 0 or 1, based on a result of comparison
between a second value obtained by adding a noise value to the
first value and a threshold. An arbitration circuit allows, when
first output signals of first neuron circuits interconnected among
the neuron circuits simultaneously change based on the weight
values, updating of only one of the first output signals of the
first neuron circuits and allows, when second output signals of
second neuron circuits not interconnected simultaneously change,
updating of the second output signals.
Inventors: |
MATSUBARA; SATOSHI;
(Machida, JP) ; TAMURA; Hirotaka; (Yokohama,
JP) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
FUJITSU LIMITED |
Kawasaki-shi |
|
JP |
|
|
Assignee: |
FUJITSU LIMITED
Kawasaki-shi
JP
|
Family ID: |
60483357 |
Appl. No.: |
15/482856 |
Filed: |
April 10, 2017 |
Current U.S.
Class: |
1/1 |
Current CPC
Class: |
G06N 3/049 20130101;
G06N 3/0445 20130101; G06N 3/063 20130101 |
International
Class: |
G06N 3/04 20060101
G06N003/04 |
Foreign Application Data
Date |
Code |
Application Number |
Jun 6, 2016 |
JP |
2016-112551 |
Claims
1. An ising device comprising: a plurality of neuron circuits which
each calculate a first value based on a sum of products each
obtained by multiplying one of a plurality of weight values each
representing connection or disconnection between a corresponding
neuron circuit and one of others of the plurality of neuron
circuits, by a corresponding one of a plurality of output signals
of the others of the plurality of neuron circuits, and output 0 or
1, based on a result of comparison between a second value obtained
by adding a noise value to the first value and a threshold; and an
arbitration circuit which allows, when a plurality of first output
signals of a plurality of first neuron circuits interconnected
among the plurality of neuron circuits simultaneously change based
on the plurality of weight values, updating of only one of the
first output signals of the plurality of first neuron circuits and
allows, when a plurality of second output signals of a plurality of
second neuron circuits not interconnected simultaneously change,
updating of the plurality of second output signals.
2. The ising device according to claim 1, wherein the arbitration
circuit generates first identification information that identifies
one of the plurality of first neuron circuits, and wherein, among
the plurality of neuron circuits, a third neuron circuit
corresponding to the first identification information performs
updating, based on the plurality of output signals.
3. The ising device according to claim 2, comprising: a memory that
holds a group of items of identification information that
respectively identifies the plurality of neuron circuits, wherein
the arbitration circuit reads the group of items of identification
information from the memory, extracts a group of items of first
identification information that respectively identifies the
plurality of first neuron circuits from the group of items of
identification information, and randomly selects the first
identification information from the group of items of first
identification information or selects first identification
information having a maximum or minimum value from the group of
items of first identification information.
4. The ising device according to claim 3, wherein the arbitration
circuit extracts second identification information that identifies
a fourth neuron circuit connected to the third neuron circuit based
on the plurality of weight values from the group of items of
identification information and excludes the second identification
information from the group of items of first identification
information.
5. A control method of an ising device, the method comprising: for
the ising device including a plurality of neuron circuits which
each calculate a first value based on a sum of products each
obtained by multiplying one of a plurality of weight values each
representing connection or disconnection between a corresponding
neuron circuit and one of others of the plurality of neuron
circuits, by a corresponding one of a plurality of output signals
of the others of the plurality of neuron circuits and output 0 or
1, based on a result of comparison between a second value obtained
by adding a noise value to the first value and a threshold, and an
arbitration circuit which allows, when a plurality of first output
signals of a plurality of first neuron circuits interconnected
among the plurality of neuron circuits simultaneously change based
on the plurality of weight values, updating of only one of the
first output signals of the plurality of first neuron circuits and
allows, when a plurality of second output signals of a plurality of
second neuron circuits not interconnected simultaneously change,
updating of the plurality of second output signals, setting, by a
control device coupled to the ising device, the plurality of weight
values; causing, by the control device, a random signal generation
circuit to randomly select the plurality of first neuron circuits
or the plurality of second neuron circuits to be enabled among the
plurality of neuron circuits; and controlling, by the control
device, an amplitude of the noise value outputted by a noise
generation circuit.
Description
CROSS-REFERENCE TO RELATED APPLICATION
[0001] This application is based upon and claims the benefit of
priority of the prior Japanese Patent Application No. 2016-112551,
filed on Jun. 6, 2016, the entire contents of which are
incorporated herein by reference.
FIELD
[0002] The embodiment discussed herein relates to an ising device
and a control method thereof.
BACKGROUND
[0003] Neumann computers are not very suitable for solving
multivariable optimization problems. However, to solve these
problems, there is a method that uses an ising device (or a
Boltzmann machine) that performs simulated annealing using an ising
objective function. Such an ising device performs calculation by
replacing a problem to be calculated with an ising model, which is
a model representing the behavior of a ferromagnetic spin.
[0004] There is a conventional technique for realizing an ising
device with hardware modules. In this technique, a plurality of
neuron circuits are operated in a parallel manner, each of which
simulates neuron functions. See, for example, the following
documents:
[0005] Japanese Laid-open Patent Publication No. 3-100857
[0006] S. Geman and D. Geman, "Stochastic Relaxation, Gibbs
Distributions, and the Bayesian Restoration of Images," IEEE
Transactions on Pattern Analysis and Machine Intelligence, PAMI-6,
pp. 721-741, 1984
[0007] It has been proved that convergence into an optimal solution
is achieved when a state (an output value) of a single neuron
circuit is updated at one time in a Boltzmann machine. However, it
has not been proved that such convergence is achieved when states
of a plurality of neuron circuits are simultaneously updated. Thus,
when a plurality of neuron circuits are operated in a parallel
manner, while improvement in the speed of the calculation
processing can be expected, the convergence into an optimal
solution could be deteriorated.
SUMMARY
[0008] According to one aspect, there is provided an ising device
including: a plurality of neuron circuits which each calculate a
first value based on a sum of products each obtained by multiplying
one of a plurality of weight values each representing connection or
disconnection between a corresponding neuron circuit and one of
others of the plurality of neuron circuits, by a corresponding one
of a plurality of output signals of the others of the plurality of
neuron circuits, and output 0 or 1, based on a result of comparison
between a second value obtained by adding a noise value to the
first value and a threshold; and an arbitration circuit which
allows, when a plurality of first output signals of a plurality of
first neuron circuits interconnected among the plurality of neuron
circuits simultaneously change based on the plurality of weight
values, updating of only one of the first output signals of the
plurality of first neuron circuits and allows, when a plurality of
second output signals of a plurality of second neuron circuits not
interconnected simultaneously change, updating of the plurality of
second output signals.
[0009] The object and advantages of the invention will be realized
and attained by means of the elements and combinations particularly
pointed out in the claims.
[0010] It is to be understood that both the foregoing general
description and the following detailed description are exemplary
and explanatory and are not restrictive of the invention.
BRIEF DESCRIPTION OF DRAWINGS
[0011] FIG. 1 illustrates an example of an ising device according
to an embodiment;
[0012] FIG. 2 illustrates an example of neuron circuits;
[0013] FIG. 3 illustrates an example of a probability
P.sub.i(h.sub.i) with which a state x.sub.i represents 1;
[0014] FIG. 4 illustrates an example of a random signal generation
circuit;
[0015] FIG. 5 illustrates an example of an arbitration circuit;
[0016] FIG. 6 is a flowchart illustrating an example of control
processing performed by the ising device;
[0017] FIG. 7 illustrates progress of simulated annealing; and
[0018] FIG. 8 is a flowchart illustrating an example of arbitration
and updating processing performed by the arbitration circuit.
DESCRIPTION OF EMBODIMENTS
[0019] An embodiment will be described below with reference to the
accompanying drawings, wherein like reference characters refer to
like elements throughout.
[0020] FIG. 1 illustrates an example of an ising device according
to an embodiment.
[0021] An ising device 1 includes a register unit 2a, a noise
generation circuit 3, a random signal generation circuit 4, a
plurality of (n) neuron circuits 5a1 to 5an (5a1, 5ak, and 5an in
FIG. 1), a control device 6, and an arbitration circuit 10. The
ising device 1 is realized by a single semiconductor integrated
circuit (a chip), for example.
[0022] The register unit 2a includes registers 2b1 to 2bn (2b1,
2b2, 2bk, and 2bn in FIG. 1). Each of the registers 2b1 to 2bn
holds a plurality of weight values, each of which represents
connection or disconnection between a corresponding one of the
neuron circuits 5a1 to 5an and one of the other neuron circuits.
For example, the register 2b1 holds weight values W.sub.11 to
W.sub.1n, each of which represents connection or disconnection
between the 1st neuron circuit 5a1 and one of the other neuron
circuits. The register 2bk holds weight values W.sub.k1 to
W.sub.kn, each of which represents connection or disconnection
between the k-th neuron circuit 5ak and one of the other neuron
circuits. The register 2bn holds weight values W.sub.n1 to
W.sub.nn, each of which represents connection or disconnection
between the n-th neuron circuit 5an and one of the other neuron
circuits.
[0023] In addition, FIG. 1 also illustrates an example of an
individual weight value W.sub.ij (1.ltoreq.i.ltoreq.n,
1.ltoreq.j.ltoreq.n) in a matrix. Numbers (1st, 2nd, 3rd, k-th, and
n-th in FIG. 1) of the neuron circuits 5a1 to 5an are indicated
vertically and horizontally in FIG. 1. In the example in FIG. 1, a
weight value W.sub.ii is 0, and a weight value W.sub.ji is equal to
a weight value W.sub.ij.
[0024] A weight value W.sub.ij represents connection or
disconnection between the i-th and the j-th neuron circuits of the
neuron circuits 5a1 to 5an. In addition, in FIG. 1, boxes that
correspond to a weight value (W.sub.ij.noteq.0) associated with two
neuron circuits interconnected are hatched. The other boxes that
correspond to a weight value (W.sub.ij=0) associated with two
neuron circuits not interconnected are not hatched.
[0025] For example, in the example in FIG. 1, since the weight
value W.sub.12 associated with the 1st and 2nd neuron circuits 5a1
and 5a2 is not 0, it is seen that the neuron circuits 5a1 and 5a2
are interconnected. In contrast, in the example in FIG. 1, since
the weight value W.sub.1k associated with the 1st and k-th neuron
circuits 5a1 and 5ak is 0, it is seen that the neuron circuits 5a1
and 5ak are not interconnected.
[0026] The above weight values are suitably set depending on the
problem to be calculated and are stored in the registers 2b1 to
2bn. In addition, since the weight value W.sub.ii is 0, the weight
value is not stored in the registers 2b1 to 2bn in the example in
FIG. 1. The above weight values may be stored in a semiconductor
storage device such as a flash memory or a memory such as a random
access memory (RAM), instead of being stored in the registers 2b1
to 2bn.
[0027] The noise generation circuit 3 is controlled by the control
device 6 and outputs a noise value to each of the neuron circuits
5a1 to 5an. An example of the noise value will be described
below.
[0028] The random signal generation circuit 4 is controlled by the
control device 6 and outputs a selection signal for randomly
selecting a plurality of neuron circuits to be enabled from the
neuron circuits 5a1 to 5an.
[0029] For example, the random signal generation circuit 4 may
include linear feedback shift registers (LFSRs). An example of the
random signal generation circuit 4 will be described below with
reference to FIG. 4.
[0030] Hereinafter, while an operation performed by the neuron
circuit 5a1 will briefly be described, the other neuron circuits
5a2 to 5an also perform the same operation. First, the neuron
circuit 5a1 calculates a value (which will hereinafter be referred
to as a local field value) based on a sum of products, each of
which is obtained by multiplying one of a plurality of weight
values, each of which represents connection or disconnection
between the neuron circuit 5a1 and one of the other neuron circuits
5a2 to 5an, by a corresponding one of output signals of the neuron
circuits 5a1 to 5an.
[0031] Next, the neuron circuit 5a1 adds a noise value supplied
from the noise generation circuit 3 to the local field value and
compares this value obtained by the addition with a threshold (for
example, 0). The neuron circuit 5a1 outputs 0 or 1 on the basis of
the comparison result.
[0032] An example of the neuron circuit 5a1 to 5an will be
described below with reference to FIG. 2.
[0033] The control device 6 controls the noise generation circuit 3
and the random signal generation circuit 4. For example, to realize
simulated annealing, the control device 6 gradually decreases the
amplitude of the noise value by controlling an amplifier circuit
(not illustrated) included in the noise generation circuit 3. In
addition, the control device 6 is connected to the register unit 2a
via a bus (not illustrated) and writes the above weight values in
the registers 2b1 to 2bn in the register unit 2a.
[0034] For example, the control device 6 may be realized by a
processor. Examples of the processor include a central processing
unit (CPU), a micro processing unit (MPU), a digital signal
processor (DSP), an application specific integrated circuit (ASIC),
and a programmable logic device (PLD). The processor may be a
combination of at least two of a CPU, an MPU, a DSP, an ASIC, and a
PLD. The control device 6 may be a personal computer (PC).
[0035] Of all the neuron circuits 5a1 to 5an, when a plurality of
outputs signals (states) of a plurality of neuron circuits
interconnected simultaneously change on the basis of the above
plurality of weight values, the arbitration circuit 10 allows
updating of only one of the states.
[0036] In this way, since the states of the plurality of neuron
circuits interconnected are not updated simultaneously, the
convergence into an optimal solution is improved.
[0037] As described above, by determining whether an individual
weight value is 0 or not, a plurality of neuron circuits
interconnected and a plurality of neuron circuits that are not
interconnected are determined.
[0038] In addition, of all the neuron circuits 5a1 to 5an, when the
states of a plurality of neuron circuits not interconnected
simultaneously change on the basis of the above plurality of weight
values, the arbitration circuit 10 allows updating of these
states.
[0039] In this way, since the states of the plurality of neuron
circuits that are not interconnected, namely, the states of the
neuron circuits that do not affect the convergence, are updated
simultaneously, the calculation time is shortened.
[0040] An example of the arbitration circuit 10 will be described
below with reference to FIG. 5.
[0041] (Example of Neuron Circuits)
[0042] FIG. 2 illustrates an example of neuron circuits.
[0043] More specifically, FIG. 2 illustrates an example of the
neuron circuits 5a1, 5ak, and 5an among the neuron circuits 5a1 to
5an.
[0044] The same elements between FIGS. 1 and 2 are denoted by the
same reference characters, and description thereof will be
omitted.
[0045] The neuron circuit 5a1 includes multiplier circuits 5c1l to
5c1n. The neuron circuit 5ak includes multiplier circuits 5ck1 to
5ckn. The neuron circuit 5an includes multiplier circuits 5cn1 to
5cnn. In addition, the neuron circuits 5a1 includes an adder
circuit unit 5d1, an adder circuit 5e1, and a comparator circuit
5f1. The neuron circuit 5ak also includes an adder circuit unit
5dk, an adder circuit 5ek, and a comparator circuit 5fk. The neuron
circuit 5an also includes an adder circuit unit 5dn, an adder
circuit 5en, and a comparator circuit 5fn.
[0046] The multiplier circuits 5c1l to 5c1n output products of the
weight values stored in the register 2b1 and states x.sub.1 to
x.sub.n (0 or 1) of the neuron circuits 5a1 to 5an outputted from
the arbitration circuit 10. The multiplier circuits 5ck1 to 5ckn
output products of the weight values stored in the register 2bk and
the states x.sub.1 to x.sub.n (0 or 1) of the neuron circuits 5a1
to 5an outputted from the arbitration circuit 10. The multiplier
circuits 5cn1 to 5cnn output products of the weight values stored
in the register 2bn and the states x.sub.1 to x.sub.n (0 or 1) of
the neuron circuits 5a1 to 5an outputted from the arbitration
circuit 10.
[0047] The adder circuit unit 5d1 adds up the values outputted from
the multiplier circuits 5c1l to 5c1n and outputs a sum. The adder
circuit unit 5dk adds up the values outputted from the multiplier
circuits 5ck1 to 5ckn and outputs a sum. The adder circuit unit 5dn
adds up the values outputted from the multiplier circuits 5cn1 to
5cnn and outputs a sum.
[0048] The adder circuit 5e1 adds a noise value outputted from the
noise generation circuit 3 to the value (the local field value)
outputted from the adder circuit unit 5d1 and outputs a sum. The
adder circuit 5ek adds a noise value outputted from the noise
generation circuit 3 to the value outputted from the adder circuit
unit 5dk and outputs a sum. The adder circuit 5en adds a noise
value outputted from the noise generation circuit 3 to the value
outputted from the adder circuit unit 5dn and outputs a sum.
[0049] The comparator circuit 5f1 is disabled (or enabled) by a
selection signal supplied from the random signal generation circuit
4. When the value outputted from the adder circuit 5e1 is larger
than a threshold (for example, 0), the comparator circuit 5f1
outputs 1 as the state x.sub.1 of the neuron circuit 5a1. When the
value outputted from the adder circuit 5e1 is equal to or less than
the threshold, the comparator circuit 5f1 outputs 0 as the state
x.sub.1 of the neuron circuit 5a1. The comparator circuit 5fk is
disabled (or enabled) by a selection signal supplied from the
random signal generation circuit 4. When the value outputted from
the adder circuit 5ek is larger than the threshold, the comparator
circuit 5fk outputs 1 as the state x.sub.k of the neuron circuit
5ak. When the value outputted from the adder circuit 5ek is equal
to or less than the threshold, the comparator circuit 5fk outputs 0
as the state x.sub.k of the neuron circuit 5ak. The comparator
circuit 5fn is disabled (or enabled) by a selection signal supplied
from the random signal generation circuit 4. When the value
outputted from the adder circuit 5en is larger than the threshold,
the comparator circuit 5fn outputs 1 as the state x.sub.n of the
neuron circuit 5an. When the value outputted from the adder circuit
5en is equal to or less than the threshold, the comparator circuit
5fn outputs 0 as the state x.sub.n of the neuron circuit 5an.
[0050] The neuron circuits other than the neuron circuits 5a1, 5ak,
and 5an are configured in the same way.
[0051] These neuron circuits 5a1 to 5an are hardware modules that
realize an operation of an ising energy function. For example, an
ising energy function E(x) is defined by the following expression
(1).
E ( x ) = - i , j W ij x i x j - i b i x i ( 1 ) ##EQU00001##
[0052] The first term on the right side of expression 1 is directed
to all the combinations of two neuron circuits selectable from all
the neuron circuits without missing any neuron circuit or
redundantly counting any neuron circuit. More specifically, the
first term represents the sum of products, each of which is
obtained by multiplying states of two neuron circuits by a weight
value associated with the two neuron circuits. As described above,
W.sub.ij=W.sub.ji, and W.sub.ii=0.
[0053] The second term on the right side of expression 1 represents
the sum of products, each of which is obtained by multiplying a
bias value of an individual one of the neuron circuits by a state
of the individual one of the neuron circuits. In expression 1,
b.sub.i represents the bias value of the i-th neuron circuit.
[0054] To express the above energy function E(x) by using hardware
modules, the neuron circuits 5a1 to 5an illustrated in FIG. 2
calculate local field values h.sub.1 to h.sub.n, respectively. For
example, the local field value h.sub.i of the i-th neuron circuit
is expressed by the following expression (2).
h i = j W ij x j + b i ( 2 ) ##EQU00002##
[0055] The first term on the right side of expression 2 represents
the sum of products, each of which is obtained by multiplying a
weight value representing connection or disconnection between the
i-th neuron circuit and another one of the neuron circuits by a
state of that neuron circuit.
[0056] When i=1, the local field value h.sub.i corresponds to the
value outputted from the adder circuit unit 5d1. While the bias
value is 0 in the example in FIG. 2, an adder circuit that adds a
bias value to the value outputted from the adder circuit unit 5d1
may be arranged. The bias value is supplied from the control device
6, depending on the problem.
[0057] In the neuron circuits 5a1 to 5an as illustrated in FIG. 2,
to perform simulated annealing, the comparator circuits 5f1 to 5fn
compare the values obtained by adding noise values to the local
field values h.sub.1 to h.sub.n with the thresholds,
respectively.
[0058] For example, the noise value is added so that a probability
with which an individual one of the output values of the comparator
circuits 5f1 to 5fn (the states x.sub.1 to x.sub.n of the neuron
circuits 5a1 to 5an) represents 1 accords with a sigmoid function.
For example, the noise value is added so that a probability
P.sub.i(h.sub.i) with which the state x.sub.i of the i-th neuron
circuit represents 1 satisfies the relationship of the following
expression (3).
P.sub.i(h.sub.i)=1/[1+exp(-h.sub.i/T)] (3)
[0059] In expression (3), T represents an effective
temperature.
[0060] To obtain the probability P.sub.i(h.sub.i) as illustrated in
expression (3), a probability density function p(ns) of a noise
value ns to be added is expressed by the following expression
(4).
p(ns)=exp(-h.sub.i/T)[1+exp(-h.sub.i/T)].sup.2 (4)
[0061] FIG. 3 illustrates an example of the probability
P.sub.i(h.sub.i) with which the state x.sub.i represents 1.
[0062] The horizontal axis represents a value obtained by adding
the noise value ns to the local field value h.sub.i, and the
vertical axis represents the probability with which the state
x.sub.i represents 1.
[0063] A waveform 30 indicates the probability P.sub.i(h.sub.i)
with which the state x.sub.i represents 1 when the noise value ns
that accords with the probability density function p(ns) as
illustrated in expression (4) is added to the local field value
h.sub.i. A waveform 31 indicates the probability P.sub.i(h.sub.i)
with which the state x.sub.i represents 1 when the noise value ns
is not added to the local field value h.sub.i.
[0064] As indicated by the waveform 31, if the noise value ns is
not added to the local field value h.sub.i, when the local field
value h.sub.i is equal to or less than a threshold Vth,
P.sub.i(h.sub.i)=0. However, when the local field value h.sub.i
exceeds the threshold Vth, P.sub.i(h.sub.i)=1.
[0065] In contrast, as indicated by the waveform 30, if the noise
value ns is added to the local field value h.sub.i, the probability
P.sub.i(h.sub.i) changes with the sigmoid function.
[0066] Next, an example of the random signal generation circuit 4
will be described with reference to FIG. 4.
[0067] For example, the random signal generation circuit 4 includes
LFSRs that generate a 7-bit random signal based on a primitive
polynomial of degree 7 (X.sup.7+X.sup.6+1). The random signal
generation circuit 4 includes registers 4a to 4g (flip flops), and
an XOR circuit 4h, and a decoder circuit 4i.
[0068] The registers 4a to 4g are connected in series with each
other. In addition, a signal line ("seed") for giving an initial
value and a signal line ("reset") for setting an initial value are
connected to each of the registers 4a to 4g. In addition, the
registers 4a to 4g operate in synchronization with a clock signal,
which is supplied from the control device 6, for example.
[0069] The XOR circuit 4h calculates an XOR operation on values
outputted from the registers 4f and 4g and outputs an operation
result. The XOR circuit 4h has an output terminal connected to an
input terminal of the register 4a in the initial stage.
[0070] The decoder circuit 4i decodes the 7-bit random signal
outputted from the output terminals (out0 to out6) of the registers
4a to 4g and outputs a 128-bit decode signal.
[0071] With the random signal generation circuit 4, a 7-bit random
signal is outputted from the output terminals (out0 to out6) of the
registers 4a to 4g, and the decoder circuit 4i decodes the random
signal. Consequently, an individual one of the comparator circuit
5f1 to 5fn in the neuron circuits 5a1 to 5an (n=128) is enabled or
disabled. For example, when the least significant bit of an output
terminal en represents 0, the comparator circuit 5f1 is
disabled.
[0072] The noise generation circuit 3 may also be realized by using
LFSRs.
[0073] (Example of Arbitration Circuit 10)
[0074] FIG. 5 illustrates an example of the arbitration circuit
10.
[0075] The arbitration circuit 10 includes a memory 11, a register
unit 12, a decoder circuit 13, an XOR circuit unit 14, an
identification information extraction circuit 15, a neuron
selection circuit 16, a selection circuit 17, a connected-neuron
detection circuit 18, an adder circuit 19, an identification
information detection circuit 20, and an AND circuit unit 21.
[0076] The memory 11 holds a group of items of identification
information, each of which identifies one of the neuron circuits
5a1 to 5an. For example, a semiconductor storage device such as a
flash memory or a RAM may be used as the memory 11.
[0077] The register unit 12 includes n registers in which the
states x.sub.1 to x.sub.n of the neuron circuits 5a1 to 5an are
stored. When the identification information detection circuit 20
outputs an H (High) level signal, a corresponding one of the n
registers in the register unit 12 captures only the state x.sub.i
selected by the neuron selection circuit 16 in synchronization with
a signal supplied from the AND circuit unit 21.
[0078] The decoder circuit 13 outputs a decode output signal on the
basis of identification information about a single neuron circuit
selected by the neuron selection circuit 16.
[0079] The XOR circuit unit 14 includes n XOR circuits, which
correspond to the n registers of the register unit 12. For example,
the i-th XOR circuit performs an XOR operation on input and output
values of the i-th register and outputs an operation result. In
this way, the i-th XOR circuit detects change of the state x.sub.i
of the i-th neuron circuit. For example, when the state x.sub.i of
the i-th neuron circuit has changed from the corresponding value
stored in the i-th register to a different value, the i-th XOR
circuit outputs 1. In contrast, when the state x.sub.i of the i-th
neuron circuit has not changed from the corresponding value stored
in the i-th register, the i-th XOR circuit outputs 0.
[0080] In this way, the XOR circuit unit 14 outputs values
representing change of the states x.sub.1 to x.sub.n of the neuron
circuit 5a1 to 5an, respectively.
[0081] The identification information extraction circuit 15 reads
the group of items of identification information stored in the
memory 11 in synchronization with a clock signal. In addition, on
the basis of a result outputted from the XOR circuit unit 14, the
identification information extraction circuit 15 extracts items of
identification information about neuron circuits whose states have
changed from the group of items of identification information. The
following description assumes that the states of a plurality of
neuron circuits have changed. Namely, the identification
information extraction circuit 15 extracts a group of items of
identification information about a plurality of neuron circuits
whose states have changed from the group of items of identification
information read from the memory 11.
[0082] The neuron selection circuit 16 selects identification
information about a single neuron circuit from the group of items
of identification information extracted by the identification
information extraction circuit 15. For example, from the group of
items of identification information, the neuron selection circuit
16 randomly selects identification information about a single
neuron circuit or selects identification information representing a
minimum (or maximum) value.
[0083] The selection circuit 17 selects weight values, each of
which represents connection or disconnection between the neuron
circuit corresponding to the identification information selected by
the neuron selection circuit 16 and one of the other neuron
circuits, from the plurality of weight values stored in the
register unit 2a.
[0084] The connected-neuron detection circuit 18 refers to the
weight values selected by the selection circuit 17 and determines
weight values that represent 0. Namely, the connected-neuron
detection circuit 18 detects other neuron circuits interconnected
to the neuron circuit corresponding to the identification
information selected by the neuron selection circuit 16. Next, from
the memory 11, the connected-neuron detection circuit 18 acquires
identification information about the other neuron circuits detected
and outputs the acquired identification information.
[0085] The adder circuit 19 excludes the identification information
about the neuron circuit selected by the neuron selection circuit
16 and the identification information outputted by the
connected-neuron detection circuit 18 from the group of items of
identification information outputted by the identification
information extraction circuit 15.
[0086] The identification information detection circuit 20 detects
whether any identification information has been outputted by the
adder circuit 19. For example, the identification information
detection circuit 20 outputs a signal that changes from an L (Low)
level to an H (High) level each time the identification information
detection circuit 20 detects identification information.
[0087] The AND circuit unit 21 includes n AND circuits which
correspond to the n registers of the register unit 12. For example,
the i-th AND circuit performs an AND operation on the H-level
signal outputted from the identification information detection
circuit 20 and the decode output signal based on the identification
information about the i-th neuron circuit supplied by the decoder
circuit 13 and outputs an operation result to the i-th register. In
this way, the i-th AND circuit updates the state x.sub.i of the
i-th neuron circuit.
[0088] Next, control processing performed by the above ising device
1 will be described with reference to FIG. 6. The ising device 1
performs this processing when calculating a problem.
[0089] FIG. 6 is a flowchart illustrating an example of control
processing performed by the ising device.
[0090] [Step S1] The control device 6 initializes the register unit
2a, etc.
[0091] [Step S2] The control device 6 sets weight values based on a
problem to be calculated in each of the registers 2b1 to 2bn of the
register unit 2a.
[0092] These weight values are supplied to the arbitration circuit
10.
[0093] [Step S3] The control device 6 causes the random signal
generation circuit 4 to output a selection signal, to select
(enable) a plurality of neuron circuits randomly.
[0094] To realize simulated annealing, each time the control device
6 performs the selection of neuron circuits, the control device 6
decreases the amplitude of the noise value by controlling the
amplifier circuit (not illustrated) included in the noise
generation circuit 3.
[0095] Next, simulated annealing will be described with reference
to FIG. 7.
[0096] FIG. 7 illustrates progress of simulated annealing.
[0097] The vertical axis represents energy E, and the horizontal
axis represents combinations q.sub.K of states of all the neuron
circuits. The combinations q.sub.K are represented by "000 . . . 0"
to "111 . . . 1". FIG. 7 illustrates how the convergence into an
optimal solution is achieved as the noise amplitude is decreased
from W1 to W3. Decreasing the noise amplitude means decreasing the
effective temperature T in expression (4).
[0098] When the noise amplitude is W1, even when an individual
solution reaches a local solution q.sub.k1, q.sub.k2, q.sub.k4, or
q.sub.k5 (a solution when the energy represents a local minimal
value), the local solution can change in the direction of higher
energy. Thus, these solutions can shift from the local solutions.
As the noise amplitude gradually decreases to W2 and W3, the change
of the individual solution is also gradually limited, and the
convergence into an optimal solution q.sub.k3 (a solution when the
energy represents a minimum value) is finally achieved.
[0099] [Step S4] In step S4, the arbitration circuit 10 performs
arbitration and updating processing. This processing in step S4
will be described below.
[0100] [Step S5] The control device 6 determines whether the number
of operations of outputting identification information about an
individual neuron circuit allowed to be updated by the arbitration
circuit 10 has reached a predetermined number cnt1.
[0101] If so (YES in step S5), the processing proceeds to step S6.
If not (NO in step S5), the processing returns to step S4.
[0102] [Step S6] The control device 6 determines whether the number
of operations of selecting neuron circuits in step S3 has reached a
predetermined number cnt2.
[0103] If so (YES in step S6), the control device 6 ends the
control processing. If not (NO in step S6), the processing returns
to step S3.
[0104] Next, the arbitration and updating processing performed by
the arbitration circuit 10 illustrated in FIG. 5 will be described
with reference to FIG. 8.
[0105] FIG. 8 is a flowchart illustrating an example of the
arbitration and updating processing performed by the arbitration
circuit 10.
[0106] [Step S4a] The identification information extraction circuit
15 reads the group of items of identification information from the
memory 11 in synchronization with a clock signal. In addition, on
the basis of the output result from the XOR circuit unit 14, the
identification information extraction circuit 15 extracts a group
of items of identification information including identification
information about a plurality of neuron circuits whose states have
changed from the group of items of identification information read
from the memory 11. The identification information extraction
circuit 15 outputs the extracted identification information.
[0107] [Step S4b] The neuron selection circuit 16 selects
identification information about a single neuron circuit from the
group of items of identification information extracted by the
identification information extraction circuit 15. For example, from
the group of items of identification information, the neuron
selection circuit 16 randomly selects identification information
about a single neuron circuit or selects identification information
representing a minimum (or maximum) value.
[0108] [Step S4c] When any identification information has been
outputted by the adder circuit 19, the identification information
detection circuit 20 outputs an H-level signal. The decoder circuit
13 outputs a decode output signal on the basis of the
identification information about the single neuron circuit selected
by the neuron selection circuit 16. The AND circuit unit 21
performs an AND operation on the basis of the H-level signal and
the decode output signal and outputs signals indicating the
operation result. Depending on the signals outputted by the AND
circuit unit 21, the n registers of the register unit 12 update the
state x.sub.i corresponding to the identification information
selected by the neuron selection circuit 16 and outputs the
updating result.
[0109] [Step S4d] The connected-neuron detection circuit 18 refers
to the weight values selected by the selection circuit 17 and
determines which one of the weight values represents 0. Namely, the
connected-neuron detection circuit 18 detects other neuron circuits
interconnected to the neuron circuit corresponding to the
identification information selected by the neuron selection circuit
16. Next, from the memory 11, the connected-neuron detection
circuit 18 acquires identification information about the other
neuron circuits detected and outputs the acquired identification
information.
[0110] [Step S4e] The adder circuit 19 excludes the identification
information about the neuron circuit selected by the neuron
selection circuit 16 and the identification information outputted
by the connected-neuron detection circuit 18 from the group of
items of identification information outputted by the identification
information extraction circuit 15.
[0111] [Step S4f] The identification information detection circuit
20 detects (determines) whether any identification information has
been outputted by the adder circuit 19. If any identification
information has been outputted (YES in step S4f), the processing
returns to step S4b. If not (NO in step S4f), the arbitration
circuit 10 ends the arbitration and updating processing.
[0112] In this way, among the states x.sub.1 to x.sub.n outputted
by the respective neuron circuits 5a1 to 5an, the state x.sub.i
corresponding to the identification information selected by the
neuron selection circuit 16 is updated and outputted.
[0113] As a result, since the states of a plurality of neuron
circuits interconnected are not updated simultaneously, the
convergence into an optimal solution is not deteriorated. In
addition, since the states of a plurality of neuron circuits that
are not interconnected, namely, a plurality of neuron circuits that
do not affect the convergence, are updated simultaneously, the
calculation time is shortened. In other words, even when the
plurality of neuron circuits 5a1 to 5an in the ising device 1 are
operated in a parallel manner, the calculation time is shortened
without adversely affecting the convergence into an optimal
solution.
[0114] In an optimization problem, as many as 10.sup.3 to 10.sup.6
variables could be used, and the calculation time could be extended
significantly. In particular, as the processing comes close to the
convergence into an optimal solution, since the number of neuron
circuits able to decrease the overall energy by the updating of the
states is decreased, the probability with which an effective neuron
circuit can be selected by random selection from all the neuron
circuits is decreased. Thus, operations of unnecessary neuron
circuits whose states do not change are increased, and the
calculation time is extended significantly. In practice, effective
updating is performed one in 10.sup.3 operations or less.
[0115] However, with the ising device 1 according to the present
embodiment, for example, when 1,000 neuron circuits are operated in
a parallel manner, the probability with which neuron circuits whose
states change are detected is improved 1000 times. In this way, the
probability with which the operations of unnecessary neuron
circuits whose states have not changed occur, the operations
conventionally occupying much of the calculation time, is decreased
to 1/1000. Namely, the calculation time is shortened
significantly.
[0116] An individual neuron circuit is not limited to the circuit
configuration illustrated in FIG. 2. For example, a circuit
configuration based on an algorithm called "DeGloria algorithm" may
be used. An individual neuron circuit based on the DeGloria
algorithm holds a value (a local field value) based on a sum of
values, each of which is obtained by multiplying a weight value
representing connection or disconnection between the corresponding
neuron circuit and one of the other neuron circuits by states of a
plurality of other neuron circuits. In addition, the individual
neuron circuit outputs 0 or 1 on the basis of a result of
comparison between a value obtained by adding a noise value to the
local field value and a threshold (for example, 0). In addition,
when the state of another neuron circuit changes, a neuron circuit
specified by identification information supplied by the arbitration
circuit 10 obtains a change amount of the local field value on the
basis of the changed state of the neuron circuit and updates the
local field value by adding or subtracting the change amount to or
from the local field value prior to the state change. When states
of a plurality of neuron circuits change, the arbitration circuit
10 outputs identification information about only one of a plurality
of neuron circuits interconnected. In addition, when states of a
plurality of neuron circuits change, the arbitration circuit 10
outputs identification information about a plurality of neuron
circuits not interconnected. In this way, the same advantageous
effect as that obtained by using the circuit configuration
illustrated in FIG. 2 for the neuron circuits is obtained.
[0117] While an ising device and a control method thereof according
to one aspect have been described on the basis of an embodiment,
the above description represents only an example. The present
technique is not limited to the above description.
[0118] According to one aspect, the convergence into an optimal
solution is improved.
[0119] All examples and conditional language provided herein are
intended for the pedagogical purposes of aiding the reader in
understanding the invention and the concepts contributed by the
inventor to further the art, and are not to be construed as
limitations to such specifically recited examples and conditions,
nor does the organization of such examples in the specification
relate to a showing of the superiority and inferiority of the
invention. Although one or more embodiments of the present
invention have been described in detail, it should be understood
that various changes, substitutions, and alterations could be made
hereto without departing from the spirit and scope of the
invention.
* * * * *